Optimised Parameters

Recall from Section 6.3 that there is a division between globally uniform and spatially explicit parameters in our set-up according to the PFT distribution. Table 6.2 lists besides the prior values and uncertainties also the posterior values and uncertainties for all parameters.

As mentioned already in Section 6.2.3 only a subset of the space of control variables can be constrained. In this study 17 out of the 58 dimensions could be observed from the chosen CO₂ concentration network, i.e. 17 eigen-values are larger than the absolute value of the largest

6.4 Results and Discussion 97 1980 1985 1990 1995 2000 Years -1.0 -0.5 0.0 0.5 1.0 Error [ppm] 1980 1985 1990 1995 2000 Year -1.0 -0.5 0.0 0.5 1.0 Error [ppm]

Figure 6.8: Time-series of errors (simulated − observed) in the interannual variability of concentrations for two observing sites: left panel Mauna Loa, Hawaii and right panel South Pole. Dotted lines are the averaged uncertainties over the simulation period in the observed records.

negative eigen-value of H (see Equation 6.4). This is also reflected in the reduction of the parameter uncertainties; for about 16 parameters initial uncertainties are reduced by more than 10% (among them some of the PFT specific Vmax and β parameters, the offset, the leaf fraction

of the maintenance respiration and parameters controlling the heterotrophic respiration). One of the most important parameters to derive a good fit to the data is the spatially explicit β parameter. As noted in Section 6.3.1 the β parameter controls the carbon storage efficiency of a given ecosystem. This parameter, along with NPP, defines the net carbon uptake of the terrestrial biosphere (Equation 6.23) and therefore strongly influences modelled CO₂ concentrations. Figure 6.9 shows a map of the mean β parameter weighted by the fractional coverage of the respective PFTs per gridcell and the mean weighted relative reduction in the uncertainty. It is interesting to note that the terrestrial biosphere is almost everywhere a carbon sink except the North American boreal region and a small band from Scandinavia to Central Siberia which also is mainly covered by boreal vegetation. In these regions the biosphere is a carbon source to the atmosphere. These areas also seem to be the best constraint areas by atmospheric CO₂ observations as there the uncertainty reduction for β is the highest by more than 50%. The mid latitudinal areas but also the main desert regions (Sahara, Arabian peninsula and central Australia) exhibit the highest βopt values of≈2 due to the high values for

crops, deciduous conifers and evergreen shrubs which are the dominant PFTs in these regions. The high storage capacity for crops is plausible as it is reinforced by the harvest of agricultural products. Throughout this whole area of high carbon sink capacity the uncertainty in β is almost unchanged to the initial uncertainty. Besides, the desert areas are the regions with the lowest productivity. However, the highly productive tropical areas show only a small carbon storage capacity with a βopt value of≈1.1 but are fairly well constrained (uncertainty reduced

98 First results from a prototype Carbon Cycle Data Assimilation System (CCDAS)

Figure 6.9: Map of the mean optimised β parameter weighted by the fractional coverage of the PFTs per gridcell (left panel) and accordingly the relative reduction in the uncertainty of the β parameter (right panel), white areas over land denote no reduction.

Besides the β parameter, Vmax and the ratio jtv = Jmax/Vmax are both spatially ex-

plicit parameters. For eight out of the thirteen PFTs, Vmax shows optimised values which

are more than two standard deviation higher or lower than the initial value. The Vmax,optvalue

for C₄ grass reveals the highest change from an initial value of 8 µmol(CO₂)m−2s−1 to only 0.2 µmol(CO₂)m−2s−1 suggesting a drastic reduction in the carbon assimilation of C₄ grass. Uncertainties of Vmax values for C4 grass and tropical evergreen trees are substantially reduced

whereas uncertainties for the other Vmax values are only slightly reduced. This holds even more

so for the jtv parameter with optimised values very close to the initial values and also no reduc- tion in their uncertainties. This suggests that the sensitivity of this parameter is highly colinear with the Vmax parameter which mainly reflects prior knowledge as the ratio Jmax/Vmax was

chosen as a parameter instead of Jmax directly.

Of major interest are the parameters controlling the heterotrophic respiration and here espe- cially the optimised Q₁₀values as the temperature dependency of the soil carbon decomposition has recently been under debate because of contradicting observations [Trumbore et al., 1996; Giardina and Ryan, 2000]. Unfortunately, the error in the spin-up procedure in BETHY 11 led to an unrealistic high amount of carbon in the fast soil carbon pool at the beginning of the simulation period which was also not in equilibrium with the plants productivity after the spin-up. Therfore the optimisation procedure had to chose parameter values which compensate for this artefact as can be seen by the somewhat unexpected high litter fraction transformed into slow decomposing soil carbon, fS,opt = 0.7, the relatively high Q_10,f,opt value of 2.7 suggesting

high litter decomposition rates at already moderate temperatures and a low Q_10,s,optvalue of 1.2 reducing the decomposition rates of the slow soil carbon pool with temperature. The optimised turnover time of the litter pool has a relatively large value, τf,opt = 3.2, which partly counteracts

against the fast decomposition of the oversized litter pool at the beginning of the simulation period.

Of course this error in the BETHY 11 version finally affects all parameters, therefore, no final conclusions should be drawn from their reported optimised values here. A parameter

6.4 Results and Discussion 99

Figure 6.10: Time-series of global monthly fluxes prognosed from CCDAS smoothed with a five month running mean filter.

optimisation using the corrected BETHY 12 version is currently running and showing some very different parameter values, however, this optimisation has not found a cost function minimum yet and still exhibits a large gradient to the current cost function value.